Re: [Sbcl-devel] Type derivation of literals

[Nikodemus S. <nik...@ra...>]

2009/4/22 Nikodemus Siivola <nik...@ra...>:

> The reason why this is currently a bad idea is that type inference
> doesn't deal with compound types correctly: the inoccuous THE below
> makes SBCL think that X is still a (CONS FIXNUM FIXNUM) even after the
> SETF CAR.
>
> (defun foo (y)
>  (declare (optimize safety))
>  (let ((x (the (cons fixnum fixnum) (cons 1 2))))
>    (setf (car x) y)
>    (+ (car x) (cdr x))))
>
> (foo "foo") ; => 74339983
>
> Oops... Here's the bug:
>
>  https://bugs.launchpad.net/sbcl/+bug/309445

I did some more looking around this issue.

One way to fix this would be to generate some sort of a dummy SET node
for each function that messes up a compound object (%RPLACA, %RPLACD,
destructive list functions, etc.) Seems like a principled way to fix
this -- but requires making a comprehensive list of things which can
mess things up. With the additional SET nodes -- and some additional
magic -- PROPAGATE-FROM-SETS should be able to deal with this.

Cheers,

 -- Nikodemus